What the documents look like
Weighing and sorting happen inside the warehouse fulfilment step, where weight and specification are the operating facts. The material comes from WMS, electronic scales, the order system and quality-inspection photos. Day-to-day documents include sorting task sheets, weighing labels, short-weight compensation forms and quality-inspection records. These materials update per order or per day, so the specification, weighing result and compensation status for one product can change from one task sheet to the next.
Product-data and selection questions ask how a product should be weighed to specification, how its label should be applied, and how a short weight should be compensated. Retrieval references should keep the specific document, field and product clue so each item can be checked. Quality-inspection records and compensation forms also carry a processing status. Whether a retrieved item is in a reviewed or compensated state has to be confirmed downstream; retrieval itself does not judge status.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. During implementation, verify the vector-model identifier of each base first, exclude those that do not match, and record what is observed. Do not run cross-model retrieval to compare missed recalls.
Search mode is compared across semantic, full-text and hybrid retrieval for the test question. No single mode is assumed to suit every material. The reference limit is measured in tokens; test with real materials of different lengths to see whether content stays complete, then adjust the value. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimisation handles coreference resolution and question expansion; compare the original question with the optimised question and check whether the subject was altered. Hybrid retrieval merges multiple result paths through RRF; this is a fusion process, not a switch to be turned on. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and is enabled according to settings. Retrieval output is an array and its length can be 0; the downstream path still runs when nothing is found. Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for associated bases does not extend to external systems.
Hypothetical paper exercise sample: the source material reads, "Sorting task sheet A-0715, product is loose apples, specification 500g±20g, weighing time morning shift 15 July, weigher has signed, status awaiting review; short-weight compensation form B-0715 corresponds to the same task sheet, compensation quantity 1 unit, status awaiting review." The draft sentence to be judged is, "A-0715 has completed compensation and been dispatched." The subject, object and status needed for the judgment are all present in the source: the task sheet is awaiting review, and the compensation form is awaiting review. The draft sentence states completed compensation and dispatch; the source does not record a dispatch field, and the compensation status is awaiting review. Awaiting review does not amount to completed compensation. This judgment keeps the evidence gap open and does not supply an authorisation or effective clause.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the sorting material bases whose vector-model identifier matches | Verify each base's vector-model identifier first, exclude mismatches, record what is observed |
| Search mode | Start from semantic retrieval; also try full-text retrieval on identifier-type questions | Compare hit content across modes for the test question; do not assume one mode suits every material |
| Reference limit | Start at a token amount covering one task sheet plus one compensation form | Test with real materials of different lengths for completeness, compare against incomplete cases, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then tune the threshold | This value takes effect only under semantic retrieval or result reranking; compare against labelled samples |
| Question optimisation | Enable for multi-turn follow-ups and short questions; compare original and optimised question | Check whether the subject was altered; can be turned off when the question is clear or speed matters |
| Reranking | Use when a rerank model is available and enabled per settings | Reranking performs secondary ordering and its score can combine with a threshold; compare against unreranked candidates |
Three common errors
- Treating a retrieval hit as an authorisation or a status confirmation. The input is a compensation form whose status is awaiting review. If downstream processing treats it as "already compensated", the error is taking a retrieval result as a business status. The check compares the status field in the retrieved text with the statement in the draft sentence and observes whether the two agree.
- Associating knowledge bases that use different vector models. The input is two bases whose vector-model identifiers do not match. The error is ignoring the node requirement that associated bases share the same vector model. The check verifies each base identifier, excludes mismatches, and records what is observed, without asserting what a cross-model run would produce.
- Treating RRF as a separate switch that can be turned on. The input is a hybrid-retrieval configuration. The error is looking for an RRF switch that does not exist. The check confirms the fusion process of hybrid retrieval, compares the merged result of the semantic and full-text candidate paths, and observes the ordering change.
How to verify the setup
- Verification one: input the hypothetical sample source text and the draft sentence "A-0715 has completed compensation and been dispatched". Compare the compensation status in the draft sentence against the source status "awaiting review", and compare the dispatch field against the source, which does not record one. The condition is met when the draft sentence is judged unsupported, the evidence gap is kept open, and it is not judged as completed compensation.
- Verification two: input two knowledge-base identifiers whose vector-model identifiers do not match. Compare against the node requirement that associated bases share the same vector model. The condition is met when the mismatched base is excluded and recorded, cross-model retrieval is not run to compare missed recalls, and no run result is asserted.
- Verification three: input a clearly defined paper empty-result array []. Compare against the downstream path design. The condition is met when the empty-result path is confirmed to still run, without claiming that a fictional-rule retrieval must return [], and without guaranteeing a first hit or a full recall.